Coherent High-Speed Acoustic Communication Inside a Commercial Atlantic Salmon Pen
Email address: jacob.hiden.rudander@kongsberg.com (Jacob Rudander)
Coherent High-Speed Acoustic Communication Inside a Commercial Atlantic Salmon Pen
Jacob Rudandera,c,, Thor Husøya, Paul van Walreeb, Pål Ortena,c
aKongsberg Maritime, Horten Norway
bNorwegian Defence Research Establishment, Horten Norway
cUniversity of Oslo, Oslo Norway
Abstract
Fish telemetry is a potential method to monitor the well-being of farmed sh, where a reference group of sh are equipped with sensors tags that oversee their health and wirelessly transmit the data. Many sh must be tagged to get statistically reliable information. In this paper, very-high-frequency (250 kHz) acoustic channels are characterized using data recorded in an Atlantic salmon pen containing 180 000 sh. The results reveal vertical pen channels character- ized by a direct path and a delayed tail of uctuating multipath arrivals with an angular spread and a coherence time of order 14 ms. It is demonstrated that multichannel equalization applied to a 64-element hydrophone array enables high-speed (78 kbps) coherent communication in short bursts. We discuss the possibilities of using high-speed acoustic communication for sh telemetry and show that it has the potential of serving hundreds of distributed unsynchronized sh tags.
Keywords: Fish telemetry, aquaculture, high-speed communication, multichannel equalization, very-high-frequency, acoustic communication.
1. Introduction
Aquaculture is the fastest growing animal-based food producing sector in the world [1], and in 2016 46 % of the global sh (e.g. nsh, crustaceans, molluscs,
Corresponding author
Email address: jacob.hiden.rudander@kongsberg.com (Jacob Rudander)
. . . ) production came from aquaculture. This corresponds to 80 Mt, whereof the production of nsh alone stood for 54 Mt [1].
The most intensively farmed nsh in sea-based aquaculture is the Atlantic salmon, of which 2.2 Mt was produced in 2016 [1]. Smolt sh weighing 200 g are transferred from land-based facilities to sea-based net pens, where they stay and grow for another two years before they have reached their nal weight of about 5 kg.
The most commonly used pens are of either rectangular or circular design with a circumference of up to 157 m and depth of 1050 meters. There are typically several of such pens at each location, arranged in an array inside a fjord or calm bay, although an emerging approach is oshore pens [2], also known as open ocean aquaculture, using more robust or submerged pens. Stocking densities up to 25 kg{m3 are common, which at the larger sites correspond to 200 000 Atlantic salmon per pen, or several millions of sh per site.
Keeping track of such a large number of animals is dicult, but necessary in order to respond to changes in the population or aquatic environment. Often such monitoring is required by legislation. Today most of the sh well-being [3] observations are done remotely by manual observations through cameras and echo sounders inside the pen or via on-site observation. The aquatic en- vironment and the vast number of sh make this not only a dicult but also time-consuming task. The monitoring challenge will further increase with larger and less accessible, submerged, or open ocean pens. Another limitation with current observation methods is that they do not measure relevant quantities directly, but instead infer the conditions from behavioral and visible aspects.
With the ongoing work of precision sh farming [4], attempts are being made to apply control-engineering principles to sh farms. However, in order to con- trol the biological processes with control engineering, access to relevant animal variables is required. This can be achieved by applying automated monitor- ing to sh farms, where sh telemetry is considered as an important enabling technology [4].
In sh telemetry [5, 6] individual sh are equipped with sensor tags, able
to measure and transmit properties of the environment near the sh, e.g. their position [e.g. 7, 8], depth [e.g. 9], swimming activity [e.g. 10, 9] or physiological factors as cardiac rhythm [e.g. 11]. Fish telemetry will, therefore, give valuable real-time information of the shes' current status, compared with the indirect measurements based on behavioral observations provided by cameras and echo sounders. Provided that a sucient number of representative sh are tagged, sh telemetry will be able to provide a statistically reliable assessment for the whole population.
Fish telemetry is not a new technology and is widely used in wild-sh re- search, for instance [11]. There have also been experiments with sh telemetry in sh farms [e.g. 7, 12, 9, 13, 14], where typically a standard o-the-shelf wild- sh system is used. However, the tag requirements for wild-sh research are very dierent from aquaculture monitoring. As pointed out by [4], the main limitations with available systems are 1) the complexity of tagging sh as this requires capturing the sh and surgery and 2) the poor communication perfor- mance. The present study focused on the second of these challenges.
Due to the strong attenuation of electromagnetic waves in seawater, acoustic communication is the preferred method for wireless underwater telemetry. The low sound speed of 1500 m/s can result in a large multipath delay spread and severe Doppler eects, [15]. In addition, there is strong attenuation at high frequencies [16] and high levels of ambient noise at low frequencies [17]. It is also interesting to note that even if the bandwidth in acoustic communication is limited to some tens of kilohertz, the acoustic channel is eectively an ultra- wideband channel since the carrier frequency is comparable to the bandwidth.
In sh pens, there will be additional scattering eects due to the large moving biomass, and complete blocking of the acoustic signal may occur [9], in addition to platform movement induced by surface waves. Furthermore, the sh pen is a busy location with enhanced local ambient noise, at least in the low and mid-frequency range [18].
Available acoustic sh telemetry systems are typically based on single-input and multiple-output communication and exist with a variety of modulation
schemes and frequencies depending on the intended application. Fish tags can be pure ID tags or sensor tags that, in addition to ID transmit readings from attached sensors. An example of the use of pure ID tags is [7]. Here a sh localization system based on triangulation of short phase-coded pulses was used to track up to nine cod simultaneously inside a submerged Atlantic cod pen.
The repetition time was between 1.7 3.3 s using a 307.2 kHz carrier. Although 84.91.2 % of the transmissions were successful in this study, the actual data rate was not mentioned. An example of a system using sensor tags is [12].
Fish-tag ID and depth information were transferred from two pens with 12, and 15 tagged Atlantic salmon, respectively. The transmission was repeated once every 90 s, and about 50 % of the transmissions were lost due to collisions, whereas 510 % was lost due to channel conditions. These tags took 4 s per transmission at a rate of a few bits per second, using a 69-kHz carrier and pulse- position modulation. A similar system from the same manufacturer was used in [14] to transfer depth and acceleration sensor data in separate messages from 21 tagged Atlantic salmon. In order to reduce collisions between tags, two separate frequency bands were used in the 6971 kHz range. The repetition time was 220-380 s.
Monitoring sh eectively in pens requires a new system design to increase communication performance. Coherent acoustic communication has a rich his- tory for other subsea applications [19] and allows for higher spectral eciency, but requires compensation for intersymbol interference (ISI). A commonly used method to combat ISI in underwater communication is decision-feedback equal- ization with explicit phase tracking [20]. A single-channel equalizer is a powerful tool for mitigating channel distortions but has problems in rich scattering en- vironments due to the rapid fading of channel taps. A hydrophone array can mitigate this problem. Multichannel equalizers can exploit the spatial diversity of the signal and thereby reduce ISI and fading of channel taps. Multichannel equalizers have been used with good results in a variety of underwater acoustic communication channels and frequency bands [e.g. 20].
The present paper demonstrates that is possible to meet the call for higher
bit rates, and thereby support a large number of tagged sh, by using coherent very high frequency1(VHF) communication in combination with array process- ing. Although multichannel equalizers improve communication performance by exploiting the spatial diversity, in addition to equalizing the channel, this is not enough. A sh telemetry system intended for aquaculture must be able to handle a large number of users.
The objectives of this paper are i) to characterize signal dispersion in time, frequency, and angle for a wideband VHF communication system with a center frequency of 250 kHz; ii) to determine the communication range and perfor- mance of a coherent high-speed scheme; iii) discuss the prospect of equipping a large number of sh with telemetry tags. The data in this paper come from measurements over a vertical link inside an Atlantic salmon net pen in August 2018.
The remainder of this paper is organized as follows. Section 2 describes the experiment. Channel properties for the Atlantic salmon pen acoustic channel are examined in section 3. The communication results are presented in section 4. The possibility to retrieve telemetry from a large number of tags and power constraints are discussed in section 4.4. Section 5 discusses the communication under other environmental conditions, conclusions are nally drawn in section 6.
2. Experimental setup
The experimental pen was one of four cylindrical net pens with a circum- ference of 157 m and a maximum depth of 40 m at the site (Fig. 1a). Each pen hosted 180 000 salmon with an average mass of 1.6 kg at the time of the experiment. If the sh were to be evenly distributed through the pen, this would correspond to approximately two individuals per cubic meter. However, the salmon tend to group together at higher densities depending on the current
1In this paper: low frequency < 1 kHz, midfrequency 110 kHz, high frequency 10100 kHz, and very high frequency > 100 kHz.
external and internal environmental conditions [21]. Two factors that tend to concentrate sh in the upper layers of the pen are ongoing feeding [22, 23], no- tably early in the feeding period, and light conditions (summer months) [24].
Horizontally, the salmon tend to swim in a doughnut-shaped school during day- time [23, 24], avoiding the center of the pen.
In our case, the sh were being fed throughout the whole experiment and showed a high surface activity. Direct quantitative measurements of the sh density were not made, but visual inspection indicated that sh density was higher near the pen edge. The swimming speed of visible sh was estimated from video recordings at approximately one body length per second, with a preferred counter-clockwise swimming direction. The weather was calm with 0.20.3 m waves outside the pen and somewhat less inside. Fig. 1b shows a photo from the measurement site, where the receiver and transmitter were suspended from the red buoy. The buoy was moving back and forth with the surface waves at a frequency of about 0.5 hertz.
Details of the set-up can be seen in Fig. 2. It consisted of a single trans- mitter with a 30 opening angle, which was suspended in an adjustable rope under the receiver array. The receiver, 64 hydrophones arranged in a line array at a uniform spacing λ{2(λ0.006m at 1500 m/s), was hanging 1 m under the buoy. The rear side of the receiver hydrophones was attached to a bae structure, shielding the hydrophones from direct surface reections. The mea- surement was conducted 5 m from the edge of the pen with Tx-Rx distances of 3, 8, 13 and 18 m.
The transmitted waveforms were binary phase-shift keying (BPSK) mod- ulated, shaped with a root-raised-cosine lter with a 100 % roll-o, using a bit rate of 78.125 kbps and a carrier frequency of 250 kHz. A sequence of 4 waveforms was transmitted, consisting of 3 probe signals for channel sounding, and one pseudonoise communication sequence. The probe signals were cyclic maximal length (m-)sequences of length 255, 511 and 1023 bits. All waveforms had a duration of 4 s. They were transmitted with 1-s guard interval, making the whole sequence repetitive with a 20-s period. This transmit sequence was
(a) (b)
Figure 1: a) Aerial photo of the location, the pen in the upper right was used for the measure- ment. The pipes visible in and around the pens was used for pellet transportation. (Photo courtesy of Norge i bilder, Kartverket, Geovekst og kommunene.) b) Photo of the measure- ment site, the transmitter-receiver pair was suspended from the red buoy. A jumping salmon is noticed in the upper left of the photo.
repeated for at least 15 minutes at every Tx depth. The source level of the transmitter was approximately 178 dB re 1 µPa2m2.
3. The Atlantic Salmon pen acoustic channel
This section highlights characteristics of the VHF Atlantic salmon pen acous- tic channel. The basis for the discussion is the time varying impulse response, hpt,τq, estimated through correlative channel sounding.
The acoustic channel is often modeled as a time-variant linear lter. Let zptq represent the transmitted complex baseband signal sampled at instants t. The received signalyptq is a distorted version ofzptq
yptq hpt,τq fzpτq eptq, (1) where the acoustic channel eects are captured by hpt,τq, and whereeptq rep- resents noise.
Transmission of dedicated probe signals allows to estimate the impulse re- sponse. If the channel is linear and slowly time variant with respect to the m-sequence repetition period, the estimate of the channel impulse response ˆh
50.9 m
Rope Bouy Fish feeder
Hydrophone array Buoy
Transmitter 3-18 m Rope
Surface
1 m
Figure 2: Set-up used for the measurements. The left sketch depicts the sh cage from above, and the right sketch illustrates the rope arrangement in which the transmitter and hydrophone array were suspended.
is the true channel, hpt,τq convolved with the autocorrelation function of the channel probezpτq
ˆhpt,τq yptq fzpτq rhpt,τq fzpτqs fzpτq. (2) The closer the autocorrelation zpτq fzpτq of the transmitted signal is to a true impulse, the better the estimate of the channel will be. Using repetitive channel probes, an estimate of the channel impulseˆhpt,τq, will be available for tn nT, where T is the probe repetition period and n P r0,1,2, . . . N1s.
The repetition rateT1 has to be selected with care to minimize measurement errors. Temporal aliasing occurs whenT is shorter than the delay spread, and frequency aliasing occurs when T1 is smaller than the Doppler spread. m- sequence lengths of 255, 511 and 1023 bits were used, allowing measurements of Doppler spread up to 153, 76 and 38 Hz, and delay spreads up to 3, 6 and 13 ms, respectively.
An angularly resolved version of the impulse response is obtained by beam- forming the responses obtained for the individual array elements. For a line array, using a standard delay-and-sum beamformer, the angularly resolved im-
pulse response estimate is ˆhpt,τ,θq
¸M m1
ˆhmpt,τ∆mqexppi2π∆mfcq, (3) whereM is the number of hydrophones in the array, ∆mxmsinpθq{c with xmdenoting the hydrophone position in the array, andθ is the angle of arrival.
For short ranges the element delay ∆m has to be compensated for spherical spreading of the wavefront.
We will rst provide the channel characteristics for a single element before proceeding to the angular characteristics. In the following single-element anal- ysis element analysis, the center hydrophone (m32) of the array was used.
3.1. 1-D density functions The power delay prole isE
|ˆh32pt,τq|2
, and is estimated by averaging over the observation periodN T
Pˆ32pτq 1 N T
»pN1qT 0
ˆh32pt,τq2dt . (4) It is shown in Fig. 3a for the four vertical ranges. The curves are normalized by the value of the strongest arrival at the 3-m depth. The general structure of the delay prole is a direct arrival followed by a rich tail of scattered energy that lasts about 2 ms. As the depth increases, the direct arrival becomes weaker compared with the tail. Apart from the shortest range, most of the energy is received via scattered paths. At all depths, there is a bump in the delay prole between 4 and 8 ms. Delayed paths have traveled a longer distance. The delay between the direct path and the onset of the bump corresponds to an excess path length of 7.5 m at 3-m range, which decreases to 6 m at 18-m range.
Geometric calculations show that this excess path length can be explained by reection o the oating collar surrounding the pen. The excess path length is reduced as the angle from the transmitter towards the collar decreases, which occurs for a larger depth of the transmitter. This collar is a strong reector that yields a measurable response even when the collar is outside the main lobe of the transmitter. The power carried by the bump is small compared to that of
(a) (b)
Figure 3: Power delay prole. (a) full range, (b) zoomed in on the rst 2 ms.
(a) (b)
Figure 4: (a) Doppler power spectrum and (b) corresponding autocorrelation function.
the rst 2 ms and is of little signicance for communication performance. Fig.
3b zooms in on the start of the impulse response, highlighting the tail whose length increases with the depth up to 13 m depth.
The Doppler power spectrum is estimated as Pˆ32pvq
»T 0
»pN1qT 0
ˆh32pt,τqexpp2πivtqdt
2
dτ (5)
and shown in Fig. 4a. The width of the spectrum increases with depth, as the fraction of energy received from moving scatterers in the water column increases.
The main scatterers in this environment are most likely sh and food particles.
An important design parameter for coherent communication systems is the coherence time, which indicates at what rate the channel estimates have to be
refreshed in order to track uctuations. The temporal correlation, here com- puted as the inverse Fourier transform of the Doppler power spectrum, is plotted in Fig. 4b. Dening the coherence time as the time it takes for the normalized correlation function to drop from 1 to 0.5, a value of 14ms is found for depths beyond 3 m. In this uctuating channel, communication has to occur in very short bursts, or the channel has to be continuously tracked.
3.2. 2-D density functions
The Doppler spectrum and the delay prole are informative quantities, but being 1-D density functions they do not reveal all channel characteristics con- tained in the time-varying impulse response. This section takes a closer look, using the transmission range of 18 m as an example.
The time-varying impulse response in Fig. 5a reveals an undulating rst arrival that moves back and forth in delay with varying amplitude. This is more clearly shown in Fig. 6a, which zooms in on the rst arrival. It is followed by a diuse tail of weaker arrivals. The cycle period is 0.5 hertz and the peak-to-peak excursion in the delay is about 0.8 ms, corresponding to a path length variation of 0.12 m. There are periods of shadowing, in which the direct arrival can fade by as much as 20 dB, whereas the tail is much less aected. The periodic fading is present at all depths, occurs simultaneously on all hydrophones, and matches the period of the buoy movement.
The suspected cause of this behavior is a misalignment of the receiver caused by entangled ropes. The situation is depicted in Fig. 6b, where the receiver is mounted in a frame with the hydrophone array seen from the endre direction.
Owing to a mounting oset of the transmitter, the frame does not hang horizon- tally under the buoy, and as a result the frame will pivot around its left corner as the buoy moves up and down with the waves. The movement will be per- pendicular to the hydrophone array. The pivoting will change the path length to the transmitter, as observed in Fig. 6a, and the angle of arrival. The change in angle is up to 40 degrees, which is enough to move the transmitter out of the main beam of the hydrophones. The fading will be strongest at the largest
(a) (b)
Figure 5: (a) Impulse response and (b) delay-Doppler spread function measured at vertical distance of 18 m.
angular oset, corresponding to the longest excess path length, which matches well with what is seen in Fig. 6a. This also explains why all the hydrophones fade simultaneously, with a periodicity independent of transmitter depth. The fading is, therefore not a property of the acoustic channel, but rather an artifact of an improvised set-up.
The corresponding delay-Doppler spread function is shown in Fig. 5b. The Doppler spread is larger on delayed arrivals and is fairly evenly distributed around zero hertz. With ongoing feeding, the sh where highly mobile and could be a source of excess widening of the spectrum. There are distinct points in the scattering function, which could be reections o sh moving at dierent velocities at dierent locations.
Figure 7 shows six snapshots of the angularly resolved impulse, evenly spaced throughout the 4-s probe signal. The magnitude is normalized on the strongest arrival across all snapshots and is clipped at 15dB to highlight strong point scatterers. The 10oset of the direct arrival is caused by the misaligned set- up illustrated in Fig. 6b. The snapshots are characterized by a direct arrival and point scatterers which are sparsely distributed in the angle-delay domain. These scatterers are highly mobile and move around, appearing and disappearing be- tween successive snapshots. The direct arrival is relatively stable by comparison.
(a)
Hydrophone array Buoy
Transmitter 3-18 m Rope
Surface
1 m
(b)
Figure 6: (a) Impulse response at 18 m depth, zoomed in on the rst arrivals. Note that the fading only occurs during excess timedelays. (b) Maladjustment of the improvised set-up.
Although it wanders back and forth in delay, its angle varies only by a few de- grees. The remainder of the energy arrives via a large number of weak point scatterers, which could be second-order reections from sh or reections o particles in the water column. The sh were being fed during the measurement;
hence food particles might contribute. Angular dispersion and delay are not separable. Starting at 4 ms, the scatterers which cause the bump in Fig. 3a can be seen. These are also distinct point scatterers at various angles, presumably a reection o the oating collar scattered back from the biomass.
3.3. Spatial correlation
The spatial correlation is important for a communication system with array processing. The receiver spatial correlation is dened as
Zm,m1arg max
∆τ
1 T
»T 0
ˆhmpt,τqˆhm1pt,τ ∆τqdτ , (6) and a normalized version
Zˆmmˆ Zmmˆ
?ZmmZmˆmˆ
, (7)
which is the maximum value of the cross correlation divided by the square root of the product of the peak autocorrelation values.
(a) (b)
(c) (d)
(e) (f)
Figure 7: Snapshots of the angularly resolved impulse response at 18 meters depth, taken (a) 0 s, (b) 0.8, (c) 1.6, (d) 2.3, (e) 3.1 and (f) 3.8 s into the probe signal. The data are normalized on the strongest arrival across all snapshots and clipped at 15dB.
Figure 8: Average spatial correlation coecient at four dierent depths .
The spatial correlation averaged over the whole measurement interval of the 64 hydrophones is plotted in Fig. 8, indicating that except for very short distances, the correlation is low. Using 0.5 as a criterion for uncorrelated signals, the signal spatial coherence length is about three hydrophone distances at 8, 13, and 18 m. At 3 m, the signal never decorrelates during the measurement interval. The reason for the low spatial coherence is that, except at 3 m, most of the energy is received via scattered paths. The contribution of noise to the measured decorrelation is negligible, as the SNR of the recorded signals exceeds 50 dB at all ranges.
The low spatial correlation indicates that diversity combining solutions may have an advantage over beamforming, yielding diversity gain with every third added hydrophone. On the other hand, the high SNR may permit harvesting sucient energy from a single specular path with a beamformer, mitigating ISI by spatially ltering out diuse scattered paths.
Table 1: Experiment conditions and parameters
Depth (m) 3 8 13 18
Input SNR per hydrophone (dB) 58 55 52 51
τ90(ms) 0.14 0.69 2.18 3.2
τ50(ms) 0.03 0.13 0.12 0.12
f90 (Hz) 37 42.1 58.6 70.7
f50 (Hz) 13.8 12.8 16.4 19.0
θ90 (degrees) 66.2 97.4 96.7 96.7
θ50 (degrees) 8.5 14.9 21.3 19.5
Tc (ms) 21 15 14 13
Lc (hydrophone spacing) - 2.8 3 2.5
3.4. Summary of channel characteristics
The sh pen channel characteristics are summarized in Table 1. The depth is the vertical distance from the transmitter to the receiver. The delay (τ90
andτ50), Doppler (f90andf50) and angular spread (θ90 andθ50) are dened as the shortest intervals that contain 90 % respectively 50 % of the total received energy. The coherence time Tc and coherent lengthLc is dened as the half peak width of the normalized correlation functions rst 0.5 crossing points.
4. A receiver for short burst coherent communication
In previous sections, the VHF Atlantic salmon pen channel was presented.
It is a highly uctuating channel made up of direct arrival and strong point scat- terers causing severe ISI, which renders coherent communication a challenging task. A well-established method to cope with ISI is multichannel equalization.
Common approaches are either directly forwarding the hydrophone signals to a
multichannel equalizer, or by beamforming the hydrophone signals, forwarding one or several beams to the equalizer.
Multichannel equalizers can act as a beamformer and/or exploit spatial di- versity, thereby reducing ISI and fading of channel taps, and have been used with good results in demanding underwater acoustic communication [e.g. 20].
A beamformer can steer beams towards distinct scatterers. This has been used with improved communication performance in the mid frequency band [25], here a reduced Doppler distortion and delay spread were reported, compared to the overall channel, when pointing narrow beams towards distinct angularly resolved arrivals. The drawback of beamforming is the requirement of high spatial coherence to obtain beamforming gain. When spatial coherence is low, the actual gain of beamforming will be limited. The spatial coherence for our channel was discussed in section 3.3 and showed a low spatial coherence for the three longest transmission ranges.
Two strategies are possible to deal with the short coherence time: i) com- municate in short bursts, where the burst is shorter than the channel coherence time, or ii) use an adaptive receiver that tracks the channel. The rst strategy is attractive for sh telemetry since the sensor data, for instance, swimming depth, heart rate, acceleration, and respiratory activity, require a packet size of a few bytes. The remaining section, therefore, focuses on a short-burst communication system that assumes a static channel during transmission.
4.1. Bit sequence
At a symbol rate of 78.125 kBd the channel will be decorrelated over 1100 symbols at a typical coherence time of 14 ms (Sec 3). The typical delay-spread which carries 50 % of the energy is <0.13 ms. Using these characteristics, a bit sequence consisting of 256 bits, whereofLt128bits are training andLp128 are payload data, seems suitable and is well within the coherence time. This would give the pulse a 100 % overhead that negatively aects the eective bit rate and increases the collision probability, but it favors the channel estimate.
The communication performance is evaluated by processing the pseudoran- dom sequences as if it contained uncoded binary phase-shift keyed communica- tion frames.
4.2. Receiver
The receiver in Fig. 9 uses an ordinary least-squares estimator to estimate the inverse channel that minimizes the error between the known transmitted training sequence and the received training sequence. The use of direct equaliza- tion combined with ordinary least-squares estimators is motivated by its robust- ness and simplicity, requiring no knowledge of the noise variance and statistical properties of the channel.
In the following discussion, it is assumed that the received signal has been brought to baseband and that the Doppler shift has been compensated for by adjusting the carrier, and if necessary using standard resampling prior to the receiver. Because the burst is very short, resampling will only be necessary in special cases. Frame synchronization is achieved by match-ltering for the pseudorandom training sequence at the start of the frame. The strongest peak in the lter output is taken as the start of the frame. The sampling rate into the receiver isTs2{TbwhereTbis the bit duration. The equivalent baseband signal isymptq and thekth discrete-time sample isympkTsq (Fig. 9).
Let Yk,mˆ be the received signal synchronized on the start of training se- quence, sampleˆkat hydrophonem, arranged in matrix form
Yˆk,m
ymprˆkL{2sTsq ymprˆkL{2 1sTsq . . . ymprˆk L{21sTsq ymprˆkL{2 2sTsq ymprˆkL{2 3sTsq . . . ymprˆk L{2 1sTsq
... ... ... ...
ymprˆk LtLsTsq ymprˆk LtL 1sTsq . . . ymprkˆ Lt1sTsq ,
(8) where L 39 is the number of lter taps per hydrophone or beam and Lt is the length of the training sequence. The total number of taps is64392496 when all hydrophone channels are used. Stacking signal matrices from dierent
Beamformer Equalizer y1(t)
ym(t)
yM(t) z(t)
hM(t, τ) h1(t, τ)
hm(t, τ)
ˆ z(kTb)
Analogfrontend
y1(kTs)
ym(kTs)
yM(kT s)
Figure 9: Equivalent baseband representation of receiver block diagram illustrating the three receiver strategies where the beamformer block can be activated, duplicated or completely bypassed.
hydrophones (or beams) gives the full received training signal matrix Yˆk
Yk,1ˆ ,Yˆk,2, . . .Yk,Mˆ
. (9)
The actual size ofYkˆdepends on the number of used hydrophones or beams. In order to demodulate the payload, the set of lter coecients which minimizes the error between knownztand received training sequenceYkˆ has to be found
gLSarg,min
g kztYˆkgk2, (10) where gLS is a column vector with lter taps stacked for the dierent input channels, andztis the training bit sequence vector. The solution to (10) is
gLSYˆkzt, (11) where denotes the pseudoinverse. A regularized inversion is used to increase the stability. An estimate of the transmitted payload bit vector is then
ˆzpYpgLS (12)
whereYp has the same structure asYkˆbut contains the unknown payload part of the signal. Because of the short duration, neither adaptability nor a digital phase-locked loop showed any improvement, and were therefore omitted from the receiver.
4.3. Result and discussion
This section evaluates three dierent receiver designs, targeting sh teleme- try in short bursts. All receivers use the equalizer described in the previous section, but they dier in the preprocessing stage. The rst receiver uses a delay-and-sum beamformer to steer a narrow beam towards the strongest ar- rival. This receiver has the benet of a small and deterministic beampattern, eectively suppressing out-of-beam interference. It has a low complexity as it feeds a single beam to the equalizer, it is hereafter denoted SBE (single beam equalizer). The second receiver has no preprocessing stage but instead directly feeds 64 hydrophones to a 64-channel equalizer, denoted MCE (multichannel equalizer). It has high complexity but enjoys the full exibility of multichannel equalization. The third receiver is a hybrid combination, splitting the array into 16 subarrays, all steered in the direction of the strongest arrival. These beams serve as input to a 16-channel equalizer, denoted SAE (subarray equal- izer). By splitting the array into 16 subarrays, each newly formed array contains four hydrophones, which is in agreement with the measured spatial coherence length. While being more complex than the SBE, it has lower complexity than the MCE.
4.3.1. Comparison of receiver strategies
Figure 10 shows the average output SNR for all equalizer strategies using a dataset of 800 frames for each depth. The MCE and SAE outperform the SBE for the same dataset and the same aperture. The SBE performs at best similar to the multichannel strategies, but never better in these channels. The dierence between the SBE and multichannel strategies increases with the depth, in line with the reduction in spatial coherence seen in Fig. 8.
The MCE and SAE yield good results at all depths. As the depth increases, it may eventually be hard to maintain a useful output signal-to-noise ratio (SNR) (Ç 6 dB) due to channel degradation, unless the array size increases. It is important to note that the degradation in output SNR solely comes from ISI, as the input SNR was ¡50 dB at all distances. The approach with the SAE
Figure 10: Comparison between the SBE (using the strongest beam), the MCE, and the SAE.
results in a marginal reduction in average output SNR compared to the MCE.
4.3.2. Angular response
In Fig. 10 the SBE output SNR came from the strongest beam, which is sub- optimal. Finding the beam with the highest output SNR is not always a trivial task, as it may dier from the strongest beam. This is exemplied in Fig. 11 which shows the unbiased receiver output SNR for the SBE and includes the beampattern for comparison. The direction of the highest output SNR coincides with either the highest or the second-highest peak in the beampattern. At short range, there is one dominant direction of arrival. As the depth increases, both the beampattern and output SNR obtain more structure, with several directions that could have been used for communication. However, the peaks are narrow, rendering the performance of a single-beam solution sensitive to pointing errors if the sequence gets longer, requiring the system to track the strongest beam adaptively.
To illustrate the dierence between single and multichannel receiver strate-
(a) (b)
(c) (d)
Figure 11: Beampattern obtained with a 64-element beamformer, and corresponding output SNR of the SBE. a) 3 m b) 8 m c) 13 m d) 18 m.
(a) (b)
(c) (d)
Figure 12: Beampattern and beamformed equalizer lter taps, a) 3 m b) 8 m c) 13 m d) 18 m. Note the match at increased depth.
gies, the MCE lter coecients have been beamformed and are shown together with the beampattern in Fig. 12. At 3 m depth, there is a distinct main lobe in the same direction as the signal arrival, which indicates that the MCE coher- ently combines the channels, similar to the SBE. This explains why the average output SNRs at 3 m are close. As the depth increases and the spatial coherence reduces, the MCE beampattern widens and eventually matches the beamform- ing beampattern fairly well. This indicates that the MCE now works more like a diversity combiner where arrivals with dierent directions are coherently combined in the equalizer, unlike the SBE, which only uses one direction. This explains why the performance dierence increases with depth.
The fact that the equalizer harvests most energy from a relatively narrow angular region suggests that it can be constrained, and does not require the full exibility of 64 hydrophones with a relatively wide opening angle. Subarray beamforming, which is used in the SAE, results in beams with a smaller opening angle, with the benet of i) a reduced receiver complexity ii) a higher input SNR into the equalizer, and perhaps most importantly, iii) reducing interference from out-of-beam arrivals. In order for subarray beamforming to be successful, the spatial coherence within the subarray has to be high. In this case, the subarray length slightly exceeds the spatial coherence length. The resulting loss was, however, negligible, as shown in Fig. 10.
4.3.3. Spatial response
As shown in Fig. 10, a MCE outperforms a SBE. The reason is its ability to exploit spatial diversity, as shown in section 3, giving it a signicant advan- tage under dicult channel conditions. However, the performance of an MCE depends on the array conguration. Closely spaced hydrophones will be cor- related, reducing the potential diversity gain. If the aperture is small, it will fail to pinpoint narrow arrivals, whereas a large spacing invokes grating lobes.
The present section examines the array conguration. In doing so, we select M hydrophones from the set of 64 available hydrophones. The selection pro- cess considers either a xed aperture or a xed spacing; the latter results in an
Figure 13: MCE receiver mean output SNR as function of number of active hydrophones, using two dierent methods to populate the array. 1) Solid line: xed aperture, evenly lling the space between the end hydrophones (m1andm64); 2) dashed line: xed spacing.
aperture size that grows withM.
Figure 13 shows the communication performance in terms of output SNR improves with M, using the same 800 frames as in section 4.3.1. The xed- aperture strategy increases rapidly withM and then attens out. The gain is close to the full gain already at 20 hydrophones, which is in agreement with the measured spatial coherence length. The xed-spacing strategy results in a steady increment of output SNR that matches the xed-aperture method rst when the array is fully populated. It appears that aperture is more important than the number of hydrophones. Indeed, closely spaced hydrophones provide essentially the same information to the receiver, and possibilities for spatial ltering are limited with wide beams.
Table 2: Received frames with one or more bit errors, for a SBE, SAE and MCE.
d, depth (m) Frames received with bit errors (%)
SBE SAE MCE
3 0 0 0
8 8.6 0.75 0.12
13 26 12 12
18 43 26 27
4.3.4. Bit errors
The same 800 frames, as used in previous result, have been analysed for bit errors. The fraction of frames received with one or more bit errors are listed in Table 2. The rst two ranges are basically error free for the SAE and MCE receiver. At 13 m frames with bit-error start to appear more frequency, and at 18 m the best result is 74 % error free frames. This was for the SAE, closely followed by the MCE with 73 % error free frames. The errors were most frequent in the periods of shadowing of the direct arrival. More than half of the erroneous frames had only one or two errors, which could be easily handled with a little source coding.
4.3.5. Communication during periods of shadowing
As shown in Fig. 5a the channels are characterized by periodic fading by as much as 20 dB. This shadowing predominantly aected the rst arrival, increas- ing the relative strength of the reverberant tail. In Fig. 14 the output SNR has been calculated for short bursts (256 symbols whereof 128 symbols for train- ing), repeated every 12 ms over the 4-s duration of the pseudorandom signal.
It is clear that the output SNR is aected by the shadowing, and that the SBE is aected most, dropping by 20 dB. This indicates that the results presented in this paper might be overly pessimistic regarding the communication perfor- mance within a sh pen because the fading is a consequence of an improvised
(a) (b)
Figure 14: Instantaneous output SNR measured over 5 s shown together with hydrophone average input power at a) 3 m b) 18 m.
set-up.
4.4. Hypothetical sh telemetry system
It has been demonstrated that, with the help of a large hydrophone array, high-speed coherent burst communication is a viable approach for sh telemetry inside an Atlantic salmon pen. This section discusses the prospects of achieving real communication with a large number of tags, as called for in [4], and the tag power requirements.
4.4.1. Number of tags
Fish tags ought to be simple and lack any form of synchronization or channel sensing, simply transmitting at a xed repetition rate. This will result in col- lisions when transmissions of multiple tags are received simultaneously. While some colliding tags can be resolved this will eventually limit the performance.
To give a lower bound, this section estimates the number of collision free sh tags per pen assuming a omnidirectional receiver. The burst length is 3.2 ms (256 symbols at 78.125 kBd), and each tag transmits at a repetition rate of 0.1 Hz, this gives 3052 available time slots. Here no distinction is made between fully or partially overlapping messages, this reduces the number of time slots to 1526. The probability of multiple transmissions in the same slot can be formulated as a ball and urn problem. The probability of getting exactly k
Figure 15: Number of collision-free slots for a varying number of tags.
transmissions in a specic time slot is given by the probability mass function of the binomial distribution
PrpX kq Ntag
k
1 Nslot
k 1 1
Nslot
Ntagk
. (13)
The expected number of time slots with one transmission isPrpX 1q Nslot and is shown in Fig. 15 as a function of the number of tags and available time slots. A few hundred tags can be handled before collisions are noticed. The number of read tags will continue to grow untilNtag1500, but at this point, more than half of the tags will experience collisions. Beyond 1500 tags the throughput starts to decline. By reducing the tag transmission repetition rate, or by introducing angular sectors, the eective number of time slots increases.
4.4.2. Power requirement
There will be a strict power requirement on the sh telemetry tags. For practical reasons, the battery cannot be replaced and has to last for the whole deployment, which is up to two years. The input SNR was very high at all ranges
in our experiment (Table 1), and the performance was not limited by noise but by ISI. The source level of the transmitter was 178 dB re 1 µPa2m2. The question arises how much the transmit power can be reduced without aecting the performance. The VHF regime is characterized by thermal noise, which is white. Hence, in order to test noise resilience, white Gaussian noise was added to the recordings. The result is plotted in Fig. 16. The reference curve is the 64-hydrophone equalizer curve from Fig. 10, for which the input SNR exceeded 50 dB at all ranges. The other curves demonstrate that the input SNR can be reduced by at least 40 dB before the performance is aected (an output SNR of 10 dB is sucient for BPSK, or 6dB with some coding). This is equivalent to a transmitter source level of 138 dB re 1 µPa2m2. For an omnidirectional lossless transducer, this corresponds to0.3mW of electrical power.
At a symbol rate of 78.125 kBd, and a frame length of 256 bits, the energy per transmission is 1 µJ for an omnidirectional lossless 0.3 mW transducer, and at a repetition rate of 0.1 hertz, the total energy required for a two-year deployment is 6 J which is well within reach of existing battery technology. If we, for example, extend the range from 18 m to 100 m, the extra propagation loss is about 20 dB, assuming spherical spreading and an absorption of about 6 dB per 100 m at 250 kHz in seawater. The increment of 20 dB in source level corresponds, in our example tag, to 30 mW of electrical power, and energy consumption of 600 J for a two year deployment. The energy requirement for acoustical sh tag telemetry is not a limiting factor.
5. Discussion
During the eld test, the weather conditions were relatively benign. The experiments were performed over a vertical channel, using a receiver shielded for surface reections. Following a strong direct path, there was a diuse tail of weaker arrivals ascribed to scattering by sh and the pen construction. The main cause of bit errors was shadowing of the direct path due to receiver misalignment and motion by surface waves.
Figure 16: Output SNR as function of depth for four dierent input SNR noise levels. The reference curve is the actual measured output SNR.
An operational system will be working under dierent conditions. A big dierence is that the majority of links will be slant channels, which may invoke more multipath propagation than vertical channels in a shallow-water environ- ment. However, multipath due to surface and bottom reections is not expected in our scenario. The employed bae attenuates all sound from angles exceeding 90 (measured from the direction pointing down from the receiver), and thus blocks surface directions also in slant channels. The receiver array is usable to about 80, and can serve most of the salmon, except those very close to the surface. Bottom reections will be unimportant, because of the large excess path length and the strong absorption at these high frequencies.
An environmental condition that can challenge the proposed communica- tion system is wind-induced surface gravity waves. Although the oating col- lar attenuates waves to some extent, signicant waves can exist inside a pen.
Shadowing of the direct path can be avoided with a proper deployment, but
wave-induced instrument motion can lead to large Doppler eects. As the wind speed increases, breaking waves may infer clouds of air bubbles in the surface layer [e.g. 26]. While being most prominent in the rst few meters, there is evidence that bubbles can be drawn down to depths of 1020 m [26, 27] through Langmuir circulation. Air bubbles have a screening eect through scattering and absorption of sound. The strength of the screening depends on the bubble concentration, the bubble size distribution, and the acoustic frequency. The eect of wave motion and bubbles on the link quality is hard to predict, but could be mitigated by xing the receiver to the pen construction, at some depth.
On the other hand, this may require removal of the bae in order to receive tags from above the receiver, simultaneously inviting surface reections. Results cited by [21] suggest that sh swim deep during storms, which is an argument for a deep receiver.
Crowding is another factor that will aect the acoustic channel. Crowding is a process in the production cycle where the available size of the pen is reduced, and sh are constrained to a smaller volume. Crowding occurs in connection with loading to well boat and medical treatment. From an acoustic point of view, the higher sh density during crowding increases the density of scatterers.
Low-rate incoherent communication was demonstrated during crowding in [14], using eight unsynchronized receivers around the pen.
How high-speed coherent communication is aected by dierent operational scenarios and environmental conditions has to be further examined. The results from our initial experiment are promising, and measures can be taken to en- sure throughput also during more challenging conditions. For instance, multiple receivers could be deployed to serve dierent sectors of the pen, either horizon- tally and/or vertically separated. Although technically feasible, the practical and economical feasibility of more complex systems also has to be evaluated.
6. Conclusion
Utilizing data from a eld test in Bergen, Norway, August 2018, a VHF acoustic channel inside an Atlantic salmon pen has been characterized at dier- ent ranges. The time-varying impulse response was estimated with a correlative channel sounder and characterized in terms of Doppler spread, delay spread, angular spread, and spatial and temporal coherence. The pen channel is highly dynamic with a delay spread of several ms and a coherence time as short as 13 ms.
A high-speed coherent burst communication system for sh telemetry has been evaluated, where a 64-channel non-adaptive channel equalizer reveals a satisfactory performance for transmission ranges up to 18 m (the longest tested range). Calculations illustrate the prospect of retrieving data in real time from a large number of tagged sh. It appears technically possible and feasible to use high-speed coherent communication in the VHF band for sh telemetry inside Atlantic salmon pens.
Further research topics are the communication performance under dierent processes in the production cycle, under harsh weather conditions, dierent receiver congurations, and nally the required coverage and number of tagged sh.
Acknowledgment
We thank the sta at Lerøy Seafood and NCE Seafood Innovation for being central in acquiring the eld data for this paper. Special thanks to Harald Sveier for making the sh farm available and to Björgólfur Hávarðsson for help with ropes.
Comments by three anonymous reviewers helped to improve the manuscript on several counts.
Founding
This work was supported by the Norwegian Research Council under Project 251828.
References
[1] FOA, The State of World Fisheries and Aquaculture 2018 Meeting the sustainable development goals (2018).
URL http://www.fao.org/3/i9540en/I9540EN.pdf
[2] H. V. Bjelland, M. Føre, P. Lader, D. Kristiansen, I. M. Holmen, A. Fred- heim, E. I. Grøtli, D. E. Fathi, F. Oppedal, I. B. Utne, I. Schjølberg, Ex- posed aquaculture in norway, in: OCEANS 2015 - MTS/IEEE Washington, 2015, pp. 110.
[3] P. J. Ashley, Fish welfare: Current issues in aquaculture, Applied Animal Behaviour Science 104 (3) (2007) 199 235, sh Behaviour and Welfare.
doi:https://doi.org/10.1016/j.applanim.2006.09.001.
URL http://www.sciencedirect.com/science/article/pii/
S0168159106002954
[4] M. Føre, K. Frank, T. Norton, E. Svendsen, J. A. Alfredsen, T. Dempster, H. Eguiraun, W. Watson, A. Stahl, L. M. Sunde, C. Schellewald, K. R.
Skøien, M. O. Alver, D. Berckmans, Precision sh farming: A new framework to improve production in aquaculture, Biosystems Engineering 173 (2018) 176 193, advances in the Engineering of Sensor-based Monitoring and Management Systems for Precision Livestock Farming.
doi:https://doi.org/10.1016/j.biosystemseng.2017.10.014.
URL http://www.sciencedirect.com/science/article/pii/
S1537511017304488
[5] E. Thorstad, A. Rikardsen, A. Alp, F. Økland, The use of electronic tags in sh researchan overview of sh telemetry methods, Turkish Journal of Fisheries and Aquatic Sciences 13 (5) (2013) 881 896.
[6] S. J. Cooke, S. G. Hinch, M. Wikelski, R. D. Andrews, L. J. Kuchel, T. G. Wolcott, P. J. Butler, Biotelemetry: a mechanistic approach to ecology, Trends in Ecology & Evolution 19 (6) (2004) 334 343.
doi:https://doi.org/10.1016/j.tree.2004.04.003.
URL http://www.sciencedirect.com/science/article/pii/
S0169534704001120
[7] C. Rillahan, M. Chambers, W. H. Howell, W. H. Watson, A self-contained system for observing and quantifying the behavior of Atlantic cod, Gadus morhua, in an oshore aquaculture cage, Aquaculture 293 (1) (2009) 49 56. doi:https://doi.org/10.1016/j.aquaculture.2009.04.003.
URL http://www.sciencedirect.com/science/article/pii/
S0044848609003330
[8] J.-E. Juell, H. Westerberg, An ultrasonic telemetric system for auto- matic positioning of individual sh used to track atlantic salmon (salmo salar l.) in a sea cage, Aquacultural Engineering 12 (1) (1993) 1 18.
doi:https://doi.org/10.1016/0144-8609(93)90023-5.
URL http://www.sciencedirect.com/science/article/pii/
0144860993900235
[9] M. Føre, J. A. Alfredsen, A. Gronningsater, Development of two telemetry-based systems for monitoring the feeding behaviour of Atlantic salmon (Salmo salar L.) in aquaculture sea-cages, Com- puters and Electronics in Agriculture 76 (2) (2011) 240 251.
doi:https://doi.org/10.1016/j.compag.2011.02.003.
URL http://www.sciencedirect.com/science/article/pii/
S0168169911000536
[10] J. Kolarevic, . Aas-Hansen, . Espmark, G. Baeverfjord, B. F. Terjesen, B. Damsgård, The use of acoustic acceleration transmitter tags for monitoring of atlantic salmon swimming activity in recirculating aqua- culture systems (ras), Aquacultural Engineering 72-73 (2016) 30 39.
doi:https://doi.org/10.1016/j.aquaeng.2016.03.002.
URL http://www.sciencedirect.com/science/article/pii/
S0144860916300450
[11] I. G. Priede, A. H. Young, The ultrasonic telemetry of cardiac rhythms of wild brown trout (Salmo trutta L.) as an indicator of bio-energetics and behaviour, Journal of Fish Biology 10 (4) (1977) 299318. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1095- 8649.1977.tb04064.x, doi:10.1111/j.1095-8649.1977.tb04064.x.
URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.
1095-8649.1977.tb04064.x
[12] M. Føre, K. Frank, T. Dempster, J. Alfredsen, E. Høy, Biomonitoring using tagged sentinel sh and acoustic telemetry in commercial salmon aquaculture: A feasibility study, Aquacultural Engineering 78 (2017) 163 172. doi:https://doi.org/10.1016/j.aquaeng.2017.07.004.
URL http://www.sciencedirect.com/science/article/pii/
S0144860917300432
[13] E. Baras, J.-P. Lagardère, Fish telemetry in aquaculture: review and perspectives, Aquaculture International 3 (2) (1995) 77102.
doi:10.1007/BF00117876.
URL https://doi.org/10.1007/BF00117876
[14] M. Føre, E. Svendsen, J. Alfredsen, I. Uglem, N. Bloecher, H. Sveier, L. Sunde, K. Frank, Using acoustic telemetry to mon- itor the eects of crowding and delousing procedures on farmed atlantic salmon (salmo salar), Aquaculture 495 (2018) 757 765.
doi:https://doi.org/10.1016/j.aquaculture.2018.06.060.
URL http://www.sciencedirect.com/science/article/pii/
S0044848617324407
[15] P. A. van Walree, Propagation and Scattering Eects in Underwater Acous-
tic Communication Channels, IEEE Journal of Oceanic Engineering 38 (4) (2013) 614631. doi:10.1109/JOE.2013.2278913.
[16] R. E. Francois, G. R. Garrison, Sound absorption based on ocean measure- ments. Part II: Boric acid contribution and equation for total absorption, J. Acoust. Soc. Am. 72 (6) (1982) 18791890.
[17] J. A. Hildebrand, Anthropogenic and natural sources of ambient noise in the ocean, Marine Ecology Progress Series 34 (12) (2009) 19361956.
[18] C. Radford, M. Slater, Soundscapes in aquaculture systems, Aquaculture Environment Interactions 11 (2019) 53 62.
[19] D. B. Kilfoyle, A. B. Baggeroer, The state of the art in underwater acous- tic telemetry, IEEE Journal of Oceanic Engineering 25 (1) (2000) 427.
doi:10.1109/48.820733.
[20] M. Stojanovic, J. A. Catipovic, J. G. Proakis, Phase-coherent digital com- munications for underwater acoustic channels, IEEE Journal of Oceanic Engineering 19 (1) (1994) 100111. doi:10.1109/48.289455.
[21] F. Oppedal, T. Dempster, L. H. Stien, Environmental drivers of Atlantic salmon behaviour in sea-cages: A review, Aquaculture 311 (1) (2011) 1 18. doi:https://doi.org/10.1016/j.aquaculture.2010.11.020.
URL http://www.sciencedirect.com/science/article/pii/
S0044848610007933
[22] A. Fernö, I. Huse, J.-E. Juell, Åsmund Bjordal, Vertical distribution of Atlantic salmon (Salmo solar L.) in net pens: trade-o between surface light avoidance and food attraction, Aquaculture 132 (3) (1995) 285 296.
doi:https://doi.org/10.1016/0044-8486(94)00384-Z.
URL http://www.sciencedirect.com/science/article/pii/
004484869400384Z
[23] J.-E. Juell, A. Fernö, D. Furevik, I. Huse, Inuence of hunger level and food availability on the spatial distribution of Atlantic salmon,
Salmo salar L., in sea cages, Aquaculture Research 25 (4) (1994) 439451. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1365- 2109.1994.tb00709.x, doi:10.1111/j.1365-2109.1994.tb00709.x.
URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.
1365-2109.1994.tb00709.x
[24] F. Oppedal, J.-E. Juell, G. L. Tarranger, T. Hansen, Articial light and season aects vertical distribution and swimming behaviour of post-smolt Atlantic salmon in sea cages, Journal of Fish Biology 58 (6) (2001) 15701584. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1095- 8649.2001.tb02313.x, doi:10.1111/j.1095-8649.2001.tb02313.x.
URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.
1095-8649.2001.tb02313.x
[25] J. Li, Y. V. Zakharov, Ecient Use of Space-Time Clustering for Un- derwater Acoustic Communications, IEEE Journal of Oceanic Engineering PP (99) (2017) 111. doi:10.1109/JOE.2017.2688558.
[26] S. A. Thorpe, G. E. R. Deacon, On the clouds of bubbles formed by breaking wind-waves in deep water, and their role in air-sea gas transfer, Philosophical Transactions of the Royal Society of London. Se- ries A, Mathematical and Physical Sciences 304 (1483) (1982) 155210.
doi:10.1098/rsta.1982.0011.
[27] D. Farmer, M. Li, Patterns of bubble clouds organized by langmuir cir- culation, Journal of Physical Oceanography 25 (6) (1995) 14261440.
doi:10.1175/1520-0485(1995)025<1426:POBCOB>2.0.CO;2.
